You are on page 1of 20

European Congress on Computational Methods in Applied Sciences and Engineering

ECCOMAS 2004
P. Neittaanmäki, T. Rossi, S. Korotov, E. Oñate, J. Périaux, and D. Knörzer (eds.)
Jyväskylä, 24–28 July 2004

OPTION PRICING MODELS WITH JUMPS:


INTEGRO-DIFFERENTIAL EQUATIONS AND INVERSE
PROBLEMS.

Rama CONT, Peter TANKOV and Ekaterina VOLTCHKOVA

Centre de Mathématiques Appliquées


Ecole Polytechnique, F-91128 Palaiseau, France
e-mail: Rama.Cont@polytechnique.fr
www.cmap.polytechnique.fr/˜rama

Key words: Option pricing, Lévy process, jump process, integro-differential equations,
inverse problem, model calibration, regularization, entropy, viscosity solutions.

Abstract. Observation of sudden, large movements in the prices of financial assets has
led to the use of stochastic processes with discontinuous trajectories – jump processes –
as models for financial assets. Exponential Lévy models provide an analytically tractable
subclass of models with jumps and the flexibility in choice of the Lévy process allows to
calibrate the model to market prices of options and reproduce a wide variety of implied
volatility skews/smiles.
We discuss the characterization of prices of European and barrier options in exponential
Lévy models in terms of solutions of partial integro-differential equations (PIDEs). These
equations involve, in addition to a second-order differential operator, a non-local integral
term which requires specific treatment both at the theoretical and numerical level. The
study of regularity of option prices in such models shows that, unlike the diffusion case,
option price can exhibit lack of smoothness. The proper relation between option prices and
PIDEs is then expressed using the notion of viscosity solution. Numerical solution of the
PIDE allows efficient computation of option prices.
The identification of exponential Lévy models from option prices leads to an inverse
problem for such PIDEs. We describe a regularization method based on relative entropy
and its numerical implementation. This inversion algorithm, which allows to extract an
implied Lévy measure from a set of option prices, is illustrated by numerical examples.

1
Cont et al: Option pricing models with jumps

1 INTRODUCTION
The shortcomings of diffusion models in representing the risk related to large market
movements have led to the development of various option pricing models with jumps,
where large returns are represented as discontinuities in prices as a function of time.
Models with jumps allow for a more realistic representation of price dynamics and a
greater flexibility in modelling and have been the focus of much recent work [11].
Exponential Lévy models, where the market price of an asset is represented as the
exponential St = exp(rt + Xt ) of a Lévy process Xt , offer analytically tractable examples
of positive jump processes which are simple enough to allow a detailed study both in
terms of statistical properties and as models for risk-neutral dynamics i.e. option pricing
models. Option pricing with exponential Lévy models is discussed in [11, 17, 25, 36]. The
flexibility of choice of the Lévy process X allows to calibrate the model to market prices of
options and reproduce a wide variety of implied volatility skews/smiles [12]. The Markov
property of the price allows to express prices of European and barrier options in terms
of solutions of partial integro-differential equations (PIDEs) with a non-local integral
term which requires specific treatment both at the theoretical and numerical level [15].
Deterministic computational methods can then be used to compute option prices [14, 27].
We present here a summary of our recent work on partial integro-differential equations
(PIDEs) for option pricing in exponential Lévy models and related inverse problems aris-
ing in model calibration. Exponential Lévy models are presented in section 2. Under
some regularity conditions, the value of European and barrier options can be represented
in terms of solutions to an integro-differential equations (section 3). However, in section 4
we show that such regularity conditions fail to hold in many models used in finance. This
prompts us to take a closer look, in section 5, at the relation between option prices using
the notion of viscosity solution. The calibration of such models to market prices requires
extracting the model parameters – the Lévy measure and the diffusion coefficient – from
a set of observed option prices. We discuss this ill-posed inverse problem in section 6 and
propose a regularization scheme based on penalization by relative entropy, which enables
to construct an arbitrage–free risk–neutral exponential Lévy models compatible with a
given set of option prices. Section 7 outlines some directions for future research.

2 LEVY PROCESSES AND EXPONENTIAL LEVY MODELS


We consider here a class of discontinuous stochastic models in which the risk neutral
dynamics of the underlying asset is given by St = exp(rt+Xt ) where Xt is a Lévy process:
a process with independent, stationary increments with discontinuous trajectories.

2.1 Lévy processes: definitions


A Lévy process is a stochastic process Xt with stationary independent increments
which is continuous in probability (but may have discontinuous trajectories). Without
loss of generality we assume X0 = 0. The characteristic function of Xt has the following

2
Cont et al: Option pricing models with jumps

Lévy-Khinchin representation [34]:


 ∞
σ2z2
E[e izXt
] = exp tφ(z) φ(z) = − + iγz + (eizx − 1 − izx1|x|≤1 )ν(dx),
2 −∞

where σ ≥ 0 and γ are real constants and ν is a positive Radon measure on R − {0}
verifying
 +1 
x ν(dx) < ∞,
2
ν(dx) < ∞.
−1 |x|>1

The random process X can be interpreted as the independent superposition of a Brownian


motion with drift and an infinite superposition of independent (compensated) Poisson
processes with various jump sizes x, ν(dx) being the intensity of jumps of size x.
Ingeneral ν is not a finite measure: ν(dx) need not be finite. In the case where
λ = ν(dx) < +∞, the measure ν can be normalized to define a probability measure µ
which can now be interpreted as the distribution of jump sizes:
ν(dx)
µ(dx) = .
λ
The jumps of X are then described by a compound Poisson process with λ as jump
intensity (average
 number of jumps per unit time) and jump size distribution µ(.). More
generally, if |x|ν(dx) < ∞, the (possibly infinite) sum of jumps is absolutely convergent
with probability 1 and Xt can be represented as a pathwise sum of a Brownian motion
plus jumps:

Xt = σWt + γ0 t + ∆Xt (1)
0<s≤t

where γ0 = γ − |x|≤1 xν(dx). In this case the compensation of small jumps is not needed
and the Lévy-Khinchin representation reduces to:
 ∞
σ2z2
φ(z) = − + iγ0 z + (eizx − 1)ν(dx).
2 −∞

In the case where |x|ν(dx) = ∞ the jumps have infinite variation and small jumps need
to be compensated.
A Lévy process is a (strong) Markov process; the associated semigroup is a convolution
semigroup and its infinitesimal generator LX : f → LX f is an integro-differential operator
given by:
E[f (x + Xt )] − f (x)
LX f (x) = lim =
t→0 t 
σ2 ∂ 2f ∂f ∂f
= 2
+γ + ν(dy)[f (x + y) − f (x) − y1{|y|≤1} (x)] (2)
2 ∂x ∂x ∂x
which is well defined for f ∈ C 2 (R) with compact support.

3
Cont et al: Option pricing models with jumps

2.2 Exponential Lévy models


Let (St )t∈[0,T ] be the price of a financial asset modelled as a stochastic process on a
filtered probability space (Ω, F, Ft , P). Ft is usually taken to be the price history up to
t. Under the hypothesis of absence of arbitrage there exists a measure Q equivalent to P
under which the discounted prices of all financial assets are Q- martingales; in particular
the discounted underlying (e−rt St ) is a martingale under Q.
In exponential Lévy models, the (risk-neutral) dynamics of St under Q is represented
as the exponential of a Lévy process:

St = S0 ert+Xt .

Here Xt is a Lévy process (under Q) with characteristic triplet (σ,γ,ν), and the interest
rate r is included for ease of notation. The absence of arbitrage then imposes that Ŝt =
St e−rt = exp Xt is a martingale, which is equivalent to the following conditions on the
triplet (σ,γ,ν):
 
σ2
ν(dy)e < ∞ , γ = γ(σ, ν) = − − (ey − 1 − y1|y|≤1 )ν(dy).
y
(4)
|y|>1 2

We will assume this relation in the sequel. The infinitesimal generator LX then becomes:
 ∞
σ2 ∂ 2f ∂f ∂f
X
L f (x) = [ 2 − ] + ν(dy) [f (x + y) − f (x) − (ey − 1) (x)].
2 ∂x ∂x −∞ ∂x

We will also use the notation Yt = rt + Xt . Yt is then a strong Markov process with
infinitesimal generator
∂f
Lf = LX f + r . (5)
∂x
While in principle one can have both a non-zero diffusion component σ = 0 and an
infinite activity jump component, in practice the models encountered in the financial
literature are of two types: either we combine a non-zero diffusion part σ > 0 with a finite
activity jump process (in this case one speaks of a jump-diffusion model) or one totally
suppresses the diffusion part, in which case frequent small jumps are needed to generate
realistic trajectories: these are infinite activity pure jump models. Different exponential
Lévy models proposed in the financial modelling literature simply correspond to different
choices for the Lévy measure ν :

• Compound Poisson jumps: σ > 0 and ν is a finite measure.


(x−m)2
– Merton model [28]: Gaussian jumps. ν = √λ e− 2δ 2
δ 2π

4
Cont et al: Option pricing models with jumps

n
– Superposition of Poisson processes: ν = k=1 λk δyk where δx is a measure that
affects unit mass to point x.
– Kou model [24] : ν(x) = pα1 e−α1 x 1x>0 + (1 − p)α2 eα2 x 1x<0

• Infinite activity models: σ = 0 and ν(dx) = ∞

– Variance Gamma [26] ν(x) = A|x|−1 exp(−η± |x|)


– Tempered stable1 processes [23, 10]: ν(x) = A± |x|−(1+α) exp(−η± |x|)
– Normal inverse gaussian process [4]
– Hyperbolic and generalized hyperbolic processes [17, 18]
Ae−ax
– Meixner process [36]: ν(x) = x sinh(bx)

3 INTEGRO-DIFFERENTIAL EQUATIONS FOR OPTION PRICES


The value of an option with terminal payoff HT is obtained as a discounted conditional
expectation under the (risk-neutral) pricing measure Q: Ct = E[e−r(T −t) HT |Ft ].
For a European call or put, HT = H(ST ) with H(S) = (S − K)+ or H(S) = (K − S)+ .
From the Markov property,

Ct = C(t, S) = E[e−r(T −t) H(ST )|St = S].

Introducing the change of variable τ = T −t, x = ln(S/S0 ), and defining: h(x) = H(S0 ex )
and u(τ, x) = erτ C(T − τ, S0 ex ), then

u(τ, x) = E[h(x + Yτ )]. (6)

If h is in the domain of the generator L given by (5), then from (2) we see that u is the
solution of the Cauchy problem:
∂u
= Lu, on (0, T ] × R; u(0, x) = h(x), x ∈ R. (7)
∂τ
However in all cases of interest in finance, h is not smooth and does not belong to the
domain of L. More generally, by applying the Ito formula to u(t, Xt ) between 0 and T
one can show [7, 15, 31] that any smooth solution u ∈ C 1,2 of (7) has the probabilistic
representation (6):
Proposition 1 (Feynman–Kac representation for Lévy processes). Assume ∃a >
0 such that |x|>1 exp(a|x|)ν(dx) < ∞. If u ∈ C 1,2 is a classical solution of (7) and its
derivatives are bounded by a polynomial function of x, uniformly in t ∈ [0, T ], then u has
the probabilistic representation (6).
1
Also called ”truncated Lévy flights” in the physics literature [23, 10] or CGMY processes [8] in the
finance literature.

5
Cont et al: Option pricing models with jumps

The conditions on u and ν ensure that u(t, Xt ) can be represented as a martingale plus
a finite variation process; they can be weakened in various ways, see [7, 15, 33].
Similarly, barrier options can be represented in terms of solutions to initial-boundary
value problems. Consider for instance an up-and-out call option with maturity T , strike
K, and (upper) barrier U > S0 . The terminal payoff is given by HT = (ST − K)+ 1T <θ ,
where θ = inf{t ≥ 0 | St ≥ U } is the first moment when the barrier is crossed.
Due to the strong Markov property of Lévy processes, it is possible to express the value
of the option Ct = e−r(T −t) E[HT |Ft ] as a deterministic function of time t and current stock
value St before the barrier is crossed. Namely, for any (t, S) ∈ [0, T ]×(0, ∞) we can define

Cb (t, S) = e−r(T −t) E[H(SeYT −t )1T <θt ],

where H(S) = (S − K)+ and θt = inf{s ≥ t | SeYs−t ≥ U }, the first exit time after t.
Then, Ct = Cb (t, St )1t≤θ for all t ≤ T . As in the European case, by going to the log
variables we define

ub (τ, x) = erτ Cb (T − τ, S0 ex ). (8)

Again, if ub is smooth the Itô formula can be used to show that ub is a solution of the
following initial-boundary-value problem:
∂u
= Lu, on (0, T ] × (−∞, log(U/S0 )),
∂τ
u(0, x) = h(x), x < log(U/S0 ); u(τ, x) = 0, x ≥ log(U/S0 ).

Due to the nonlocal nature of the integral term, ”boundary” conditions have to be imposed
not only at the boundary but outside the boundary.2 Prices of down-and-out or double
barrier options verify similar PIDEs with Dirichlet boundary conditions. The correspon-
dence between the analytic and probabilistic properties discussed above is summarized in
Table 1.

4 SMOOTHNESS OF OPTION PRICES


In the case where the log-price Xt has a non-degenerate diffusion component, it is
known [7, 20] that the fundamental solution of the pricing PIDE, which correspond to the
density of Xt , is in fact a smooth C ∞ function. As a consequence, the option price u(t, x)
depends smoothly on the underlying and results such as Proposition 1 allows to use the
solution of the PIDE to compute the option price.
In the case of processes with a degenerate diffusion component, such as pure jump
models, the smoothness of the conditional expectation as a function of (t, S) does not
always hold, as the following example shows.
2
In fact they extend to any point that the process can jump to from inside the domain.

6
Cont et al: Option pricing models with jumps

Table 1: Correspondence between analytical and probabilistic properties

PIDE Lévy process Financial interpre-


tation

x Starting point of process Yt + x Log moneyness

Integro-differential oper- Infinitesimal generator


ator

Initial condition h Payoff at maturity,


expressed in log-price
H(S0 ex ) = h(x)

Solution of PIDE u(t, x) E[h(Yt + x)] Value of option with


payoff H(ST ), time to
maturity τ

Fundamental solution of Transition density ρt (x) of Yt e−r(T −t) × Gamma of


Cauchy problem call option

Zero boundary condition Stopping at first exit from b Knock-out barrier b


for x ≥ b

Green function with zero Density of stopped process e−r(T −t) × Gamma of
boundary condition for up-and-out call
x≥b

Comparison principle H ≥ 0 ⇒ E[H|Ft ] ≥ 0 Static arbitrage rela-


tions

7
Cont et al: Option pricing models with jumps

Example 1 (Variance Gamma process). The Variance Gamma process, introduced


by Madan & Milne [26], is a pure jump finite variation process with infinite activity,
popular in financial modelling. Its Lévy measure has a density given by:

1 Ax−B|x| θ θ2 + 2σ 2 /κ
ν(x) = e with A = 2 and B = . (9)
κ|x| σ σ2

The characteristic function of Xt , the Fourier transform of its distribution, is given by:

u2 σ 2 κ t
Φt (u) = (1 + − iθκu)− κ (10)
2
Φt (.) decays as |u|−2t/κ when |u| → ∞: the decay exponent increases with t. The funda-
mental solution ρ(t, x) of the PIDE therefore has a degree of regularity which increases
gradually with t: for t ∈ (pκ/2, (p + 1)κ/2), the fundamental solution ρ(t, .) is in C p−1 (R)
but not C p (R). For t < κ/2, ρ(t, .) is not even locally bounded. As a consequence, the
value of a European binary option defined by the payoff h(x) = 1x≥x0 is continuous but
not differentiable for t < κ/2.

The case of barrier options is even less regular. As the following example illustrates,
if no restriction is imposed on the Lévy process, the value of a barrier option – which is
formally the solution of the Dirichlet problem with zero boundary conditions – can even
turn out to be discontinuous at all times:

Example 2. Consider Xt = Nt1 (λ1 )−Nt2 (λ2 ) where Nti are independent Poisson processes
with jump intensities λ1 and λ2 . Let, for simplicity, r = 0. If λ2 = λ1 e then the
corresponding price process St = S0 eXt is a martingale.
Consider now a knock-out option which pays 1 at time T if St has not crossed the
barrier U > S0 before T , and 0 otherwise:

HT = 1T <θ(S0 ) ,

where θ(S) = inf {t ≥ 0 | SeXt ≥ U } is the first exit time if the process starts from S.
Let us show that the initial option value

C(0, S) = E[HT |S0 = S] = E[1T <θ(S) ]

is not continuous at S ∗ = U/e.


Let 0 < ε < U − S ∗ . By definition, θ(S ∗ + ε) ≤ θ(S ∗ − ε). Therefore,

C(0, S ∗ − ε) − C(0, S ∗ + ε) = E[1{θ(S ∗ +ε)≤T <θ(S ∗ −ε)} ]


= Q (θ(S ∗ + ε) ≤ T < θ(S ∗ − ε)) .

8
Cont et al: Option pricing models with jumps

Consider the following possibility: NT1 = 1 et NT2 = 0, that is, there was one positive and
no negative jumps. In this case, if St starts from S ∗ − ε it stays below U , while starting
from S ∗ + ε it crosses the barrier. This means that θ(S ∗ + ε) ≤ T < θ(S ∗ − ε). So,

C(0, S ∗ − ε) − C(0, S ∗ + ε) ≥ Q(NT1 = 1 & NT2 = 0) = e−λ1 T (e+1) λ1 T > 0.

Thus S → C(0, S) is discontinuous at S = S ∗ .

This example is a finite activity process without diffusion component. As noted above,
this case is not the interesting one in financial modelling. The following result shows
that in fact, in most cases of interest, the option price is a continuous function of the
underlying :

Proposition 2 (Continuity of European options ). If H satisfies the Lipschitz condi-


tion (17) then the forward value of a European option defined by f (τ, x) = E[H(S0 ex+rτ +Xτ )]
is continuous on [0, T ] × R.

For a proof, see [15]. Denote by by Cp+ ([0, T ] × R) the set of measurable functions on
[0, T ] × R with polynomial growth of degree p at +∞ and bounded on [0, T ] × R− :

ϕ ∈ Cp+ ([0, T ] × R) ⇐⇒ ∃C > 0, |ϕ(t, x)| ≤ C(1 + |x|p 1x>0 ). (11)

The pricing function can be shown to have polynomial growth at infinity if the payoff has
this property:

Proposition 3 (Polynomial growth). If H : (0, ∞) → [0, ∞) is Lipschitz: |H(S1 ) −


H(S2 )| ≤ C|S1 − S2 |, and there exists p > 0, such that:

H(S0 ex ) ≤ C1 (1 + |x|p ), (12)

then f (τ, x) = E[H(S0 ex+rτ +Xτ )] belongs to Cp+ ([0, T ] × R).

In general, one cannot hope for more than Lipschitz continuity; in particular uniform
bounds on derivatives, such as the ones required in [31], do not hold in cases of interest
in finance where the payoff function H is not smooth, as for call or put options. In these
cases, verification theorems such as the Proposition 1 do not apply and the option pricing
function should be seen as a viscosity solution of the PIDE (7).

5 OPTION PRICES AS VISCOSITY SOLUTIONS


Existence and uniqueness of (classical) solutions for the PIDEs considered above in
Sobolev / Hölder spaces have been studied in [7, 20] in the case where the diffusion
component is non-degenerate: for a Lévy process this simply means σ > 0 but more
generally these results apply to Markov processes with jumps where the diffusion coeffi-
cient is bounded away from zero. However many of the models in the financial modelling

9
Cont et al: Option pricing models with jumps

literature are pure jump models with σ = 0, for which such results are not available. In
fact, in pure jump models with finite variation Equation (7) is formally of first order in
the price variable so the effect of the jump term is more like a convection term rather
than a diffusion term. A notion of solution which yields existence and uniqueness for
such equations without requiring non-degeneracy of coefficients or a priori knowledge of
smoothness of solutions is the notion of viscosity solution, introduced by Crandall & Lions
for PDEs [16] and extended to integro-differential equations of the type considered here
in [1, 3, 32, 35, 37].3
Denote by U SC (respectively LSC) the class of upper semicontinuous (respectively
lower semicontinuous) functions u : [0, T ) × R → R. Let O = (l, u) ⊆ R be an open
interval, ∂O = {l, u} its boundary, and g ∈ Cp+ ([0, T ] × R \ O) a continuous function.
Consider the following initial-boundary value problem on [0, T ] × R:
∂u
= Lu, on (0, T ] × O, (13)
∂τ
u(0, x) = h(x), x ∈ O; u(τ, x) = g(τ, x), x∈
/ O. (14)
Definition 1 (Viscosity solution). A function u ∈ U SC is a viscosity subsolution of
(13)–(14) if for any test function ϕ ∈ C 2 ([0, T ] × R) ∩ Cp+ ([0, T ] × R) and any global
maximum point (τ, x) ∈ [0, T ] × R of u − ϕ, the following properties are verified:
 
∂ϕ
if (τ, x) ∈ (0, T ] × O, − Lϕ (τ, x) ≤ 0, (15)
∂τ
 
∂ϕ
if τ = 0, x ∈ O, min{ − Lϕ (τ, x), u(τ, x) − h(x)} ≤ 0,
∂τ
 
∂ϕ
if τ ∈ (0, T ], x ∈ ∂O, min{ − Lϕ (τ, x), u(τ, x) − g(τ, x)} ≤ 0,
∂τ
if x ∈/ O, u(τ, x) ≤ g(τ, x). (16)
A function u ∈ LSC is a viscosity supersolution of (13)–(14) if for any test function
ϕ ∈ C 2 ([0, T ] × R) ∩ Cp+ ([0, T ] × R) and any global minimum point (τ, x) ∈ [0, T ] × R of
u − ϕ, we have:
 
∂ϕ
if (τ, x) ∈ (0, T ] × O, − Lϕ (τ, x) ≥ 0,
∂τ
 
∂ϕ
if τ = 0, x ∈ O, max{ − Lϕ (τ, x), u(τ, x) − h(x)} ≥ 0,
∂τ
 
∂ϕ
if τ ∈ (0, T ], x ∈ ∂O, max{ − Lϕ (τ, x), u(τ, x) − g(τ, x)} ≥ 0,
∂τ
if x ∈/ O, u(τ, x) ≥ g(τ, x).
3
Definitions of viscosity solutions in these papers vary in the choice of test functions; we present here
a version which is suitable for option pricing applications.

10
Cont et al: Option pricing models with jumps

A function u ∈ Cp+ ([0, T ] × R) is called a viscosity solution of (13)–(14) if it is both a


subsolution and a supersolution. This function is then continuous on (0, T ] × R.

Note that the initial and boundary conditions are verified in a viscosity sense. The def-
inition also includes the case of initial value problems: O = R. Existence and uniqueness
of viscosity solutions for such parabolic integro-differential equations are discussed in [1] in
the case where ν is a finite measure and in [3] and [32] for general Lévy measures. Growth
conditions other than u ∈ Cp+ can be considered (see e.g. [1, 3]) with additional conditions
on the Lévy measure ν. The main tool for showing uniqueness is the comparison principle:
if u, v are viscosity solutions and u(0, x) ≥ v(0, x) then ∀τ ∈ [0, T ], u(τ, x) ≥ v(τ, x). This
property can be extended to subsolutions and supersolutions in the following sense:

Proposition 4 (Comparison principle for semi-continuous solutions[1, 22]). If


u ∈ U SC is a subsolution and v ∈ LSC is a supersolution of (13)–(14) then u ≤ v on
(0, T ] × R.

Proofs and extensions can be found in [1] for the case where ν is a bounded measure;
the case of a general Lévy measure has been recently treated in [22].
The following result, whose proof is given in [15] shows that, under rather general
conditions on the Lévy triplet and the payoff function, values of European and barrier
options can be expressed in terms of (viscosity) solutions of (13)–(14):

Proposition 5 (Option prices as viscosity solutions). Let the payoff function H


verify the Lipschitz condition on its domain of definition:

|H(S1 ) − H(S2 )| ≤ C|S1 − S2 |, ∀S1 , S2 ∈ (S0 el , S0 eu ) (17)

and let h(x) = H(S0 ex ) have polynomial growth at infinity. Then:

• The forward value of a European option u(τ, x) defined by (6) is the unique viscosity
solution of the Cauchy problem (7) (that is (13)–(14) with O = R).

• Let ub (τ, x) be the forward value of a knockout (single or double) barrier option
defined by (8). If ub (τ, x) is continuous then it is the unique viscosity solution of
(13)–(14) (with g ≡ 0).

The hypotheses above on the payoff function apply to put options, single-barrier knock-
out puts, double barrier knockout options and also to the log-contract. One can then re-
trieve call options by put-call parity. For barrier options with rebate, the zero boundary
condition has to be replaced by the value of the rebate, as in the case of diffusion models.
A discussion of sufficient conditions for continuity of value functions for barrier options is
given in [15].
A popular method for pricing European options in exp-Lévy models is the Fourier
method proposed by Carr & Madan [9], which is the method of choice when analytic

11
Cont et al: Option pricing models with jumps

expressions are available for the characteristic function of the Lévy process Xt . However
this method does not extend to barrier options or American options. Numerical solution
of PIDEs allows efficient pricing of European and barrier options on assets with jumps
and does not require analytic formulae for characteristic functions. Numerical methods
for PIDEs are discussed in [14, 27]. In particular the notion of viscosity solution turns
out to be convenient for analyzing the convergence of finite difference schemes, without
requiring smoothness with respect to the underlying [14].

6 MODEL CALIBRATION: AN ILL POSED INVERSE PROBLEM AND


ITS REGULARIZATION
A preliminary step in using an option pricing model is to obtain model parameters
– here the characteristic triplet of the Lévy process – from market data by calibrating
the model to market prices of (liquid) call options. The calibration problem for exp-
Lévy models consists of identifying the Lévy measure ν and the volatility σ from a set of
observations of call option prices:
Calibration Problem 1. Given the market prices of call options C0∗ (Ti , Ki ), i = 1..N at
t = 0, construct a Lévy process (Xt )t≥0 such that the discounted asset price St e−rt = exp Xt
is a martingale and the market call option prices C0∗ (Ti , Ki ) coincide with the prices of
these options computed in the exponential Lévy model driven by X:
∀i ∈ {1, .., N }, C0∗ (Ti , Ki ) = e−rTi E[(STi − Ki )+ |S0 ] = e−rTi E[(S0 erTi +XTi − Ki )+ ]. (18)
Problem 1 can be seen as a generalized moment problem for the Lévy process X, which
is typically an ill posed problem when the observations are finite and/or noisy: there
may be no solution at all or an infinite number of solutions and the dependence of the
solution(s) on option prices may be discontinuous, which results in numerical instabilities
in the calibration algorithm.
In this section, we first describe the commonly used nonlinear least squares method
for solving this problem and show its shortcomings. Next, we introduce a regularization
approach using relative entropy, discuss its properties and its numerical implementation.
This section is based on [12, 13].

6.1 Nonlinear least squares


The calibration problem 1 may have no solution, either because the model is misspeci-
fied or because the observed option prices, defined up to a bid-ask spread, do not lie within
the model range. An approach used to in many empirical studies of option pricing models
[2, 6] is to minimize the (squared) difference between model prices C σ,ν and market prices
C ∗ , summed over all liquid options available in the market:

N
∗ ∗
(σ , ν ) = arg inf (σ, ν) (σ, ν) = wi |C σ,ν (t = 0, S0 , Ti , Ki ) − C0∗ (Ti , Ki )|2 (19)
σ,ν
i=1

12
Cont et al: Option pricing models with jumps

where wi are positive weights chosen to balance the magnitude of the different terms;
a typical choice is to take wi−1 as the Black–Scholes Vega of the call option (Ti , Ki ).
Gradient–based methods are then used to perform the numerical minimization in (19).
Unfortunately this is a nonlinear, nonconvex optimization problem where gradient meth-
ods may give erroneous results. Figure 1 shows the shape of the nonlinear least squares
criterion in the case of two popular models, the Merton model (left) and the Variance
Gamma model (right), computed for DAX index options. In the Merton model we ob-
serve a continuum of minima corresponding to the difficulty of distinguishing, using the
option prices, the effect of jumps and volatility. In the Variance Gamma case we observe
two quite distinct parameter sets giving similar calibration performance. At a theoretical
Quadratic pricing error for Merton model: DAX options.
1000
B
900

800

700 5
x 10
600 2
500
1.8
400
1.6
300
1.4
200
1.2 A
100
1
0
3 0.8 0.25

0.6
2 0.2
8
7
6
1 5
4 0.15
3
0 0.17 0.16 0.15 0.14 2
0.21 0.2 0.19 0.18 1
0.23 0.22
0.24 0 0.1 σ
Jump intensity
Diffusion coefficient κ

Figure 1: Sum of squared differences between market prices (DAX options maturing in 10 weeks) and
model prices in the Merton model (left) and the variance gamma model as a function of σ and κ, the
third parameter being fixed.

level, while conditions can be derived [13] for well-posedness of solutions for the nonlinear
least square problem (19) they turn out to be rather restrictive and imply that the range
of model parameters is already quite well known a priori.

6.2 Regularization by relative entropy


One way to enforce uniqueness and stability of the solution is to inject prior information
into the problem by specifying a prior (exp-Lévy) model Q0 and add to the least-squares
criterion (22) a convex penalization term F (σ, ν) which measures the deviation of the
exp-Lévy model defined by (σ, ν) from the prior model Q0 :


N
(σ ∗ , ν ∗ ) = arg inf wi |C σ,ν (t = 0, S0 , Ti , Ki ) − C0∗ (Ti , Ki )|2 + αF (σ, ν) (20)
σ,ν
i=1

13
Cont et al: Option pricing models with jumps

Problem (20) can be understood as that of finding an exponential Lévy model satisfying
the conditions (1), which is closest in some sense – defined by F – to a prior exp-Lévy
model. The convex term αF is called a regularization term and is used to make the
problem well-posed. The regularization parameter α > 0 is chosen to ensure a tradeoff
between precision and stability [19].
A common choice for the regularization term is the relative entropy or Kullback Leibler
distance E(Q, Q0 ) of the the pricing measure Q with respect to the prior model Q0 :

Q dQ
F (σ, ν) = E(Q, Q0 ) = E ln
dQ0

In addition to being convex, relative entropy acts also as a barrier function for the posi-
tivity and absolute continuity constraints on (σ, ν). In the case of (risk-neutral) exp-Lévy
models, the relative entropy is easily computable in terms of the volatility σ and calibrated
Lévy measure ν [21, Theorem IV.4.39]:

 ∞ 2
T
E(Q, Q0 ) = H(ν) = 2 (e − 1)(ν − ν0 )(dx)
x
2σ −∞
 ∞ 
dν dν dν
+T ln +1− ν0 (dx) (21)
−∞ dν0 dν0 dν0

H(ν) is a convex functional of the Lévy measure ν, with a unique minimum at ν = ν0 .


Relative entropy plays the role of a pseudo-distance of the (risk-neutral) measure from the
prior and minimizing it corresponds to adding the least possible amount of information to
the prior in order to correctly reproduce observed option prices . The explicit expression
(21) of relative entropy in terms of the Lévy measure allows to construct an efficient
numerical method for finding the minimal entropy Lévy process, compatible with a set of
observed option prices.
The calibration problem now takes the form:

Calibration Problem 2. Given a prior exponential Lévy model Q0 with characteristics


(σ0 , ν0 ) find a Lévy measure ν which minimizes


N
J (ν) = αH(ν) + wi (C0ν (Ti , Ki ) − C0∗ (Ti , Ki ))2 (22)
i=1

where H(ν) is the relative entropy of the risk neutral measure with respect to the prior,
whose expression is given by (21).

The functional (22) consists of two parts: the relative entropy functional, which is
convex in its argument ν and the quadratic pricing error which measures the precision

14
Cont et al: Option pricing models with jumps

of calibration. The coefficient α, called regularization parameter defines the relative im-
portance of the two terms: it characterizes the trade-off between prior knowledge of the
Lévy measure and the information contained in option prices.
Regularization using relative entropy allows to obtain existence of solutions and conti-
nuity with respect to market data (input option prices) [13]:

Proposition 6 (Regularized calibration problem). Let the prior model Q0 be an


exp-Lévy process with jumps bounded from above.

• Existence: for each data set C ∗ = (C ∗ (Ti , Ki ), i = 1..N ) there exists an exponential
Lévy model Q which is a solution of the regularized calibration problem (22).

• The solution(s) depend continuously on the market prices: Let {C n }n≥1 and C ∗ be
data sets of option prices such that

C n − C ∗  → 0.
n→∞

For each n ≥ 1, let Qn be a solution of the calibration problem (22) with data C n
and prior Q0 . Then (Qn )n≥1 has a weakly converging subsequence and the limit of
every convergent subsequence of (Qn )n≥1 is a solution of calibration problem (22)
with data C ∗ and prior Q0 .

• Stability with respect to the prior: Let Pn be a sequence of probability measures


corresponding to exp-Lévy models with jumps uniformly bounded above by some
constant B > 0, weakly converging to Q0 . Let Qn be solution of (22) with prior Pn .
Then (Qn )n≥1 has a weakly converging subsequence and the limit of every weakly
converging subsequence is a solution of (22) with prior Q0 .

6.3 Numerical implementation


An important feature of the objective functional (22) is that its (directional) derivatives
can be computed explicitly [12]. This allows to use a gradient-based method to solve
the regularized optimization problem (22). The regularization parameter α in (22) is
determined using the Morozov discrepancy principle, as follows. First, we remark that
the observed option prices C ∗ = (C ∗ (Ki , Ti ), i = 1..N ) are defined up to a bid–ask spread;
this allows to define an a priori level for the quadratic pricing error:


N
0 = wi |Cibid − Ciask |2 (23)
i=1

Now let (σ, να ) be the solution of (22) for a given regularization parameter α > 0. Then
the a posteriori quadratic pricing error is given by (σ, να ), which one expects to be a
bit larger than 0 since by adding the entropy term we have sacrificed some precision in

15
Cont et al: Option pricing models with jumps

order to gain in stability. The Morozov discrepancy principle [29] consists in authorizing
a loss of precision that is of the same order as the a priori error by choosing α such that

0  (σ, να ) (24)

This equation need not be solved precisely: one needs simply to obtain the correct order
of magnitude for α, which is then substituted in (22) and solved to obtain the solution
of the regularized problem. Figure 2 illustrates the performance of the algorithm on a
0.5 5
simulated prior
calibrated true
4.5 calibrated
0.45

4
0.4
3.5

0.35
3
Implied volatility

0.3 2.5

0.25 2

1.5
0.2

1
0.15
0.5

0.1
6 7 8 9 10 11 12 13 14 0
Strike −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5

Figure 2: Left: calibrated vs simulated (true) implied volatilities for Kou model [24]. Right: double
exponential Lévy measure used to generate data set, prior (Gaussian measure) and calibrated measure.

simulated data set: a set of 20 option prices simulated from a Kou jump–diffusion model
[24] with diffusion coefficient σ = 0.1 was used and the prior model Q0 was chosen to
be a Merton model [28] with a (biased) diffusion coefficient σ = 0.105 and a Gaussian
measure for ν0 . The left figure shows the calibration performance: implied volatilities
generated from the Kou model are retrieved to within a few basis points. The right figure
shows the reconstructed Lévy measure: the positive bias in the diffusion coefficient of
the prior is compensated by a decrease in the intensity of small jumps. This examples
illustrates several points. First, it is difficult to distinguish the effect of (small) jumps
from volatility in prices of European options. Second, a small number of options – twenty
in this example – is sufficient to retrieve the Lévy measure. Third, a bias/error in the
estimation of volatility is compensated for by an opposite bias in the intensity of small
jumps, resulting in a precise fit of option prices.

7 PERSPECTIVES
The incorporation of jumps into option pricing models has led to models a more realistic
vision of financial risk. The use of these models, initially hampered by the lack of suitable
computational methods, has been substantially eased in recent years by the availability
of efficient numerical methods: for exponential Lévy models, efficient numerical methods

16
Cont et al: Option pricing models with jumps

have been developed for pricing European and exotic options and calibration of model
parameters.
A richer class of models, interesting for applications, is the class of stochastic volatility
models with jumps.4 While Fourier–based methods can still be applied for pricing Eu-
ropean options in such models, efficient numerical methods for calibration and pricing of
exotic options remain to be developed.
Another direction where many modelling and computational issues remain is multi-
asset models with jumps. While the computational finance literature has primarily focused
on one–dimensional problems, most applications concern multidimensional ones: multi-
asset options, interest rate options and portfolio optimization. We hope that some readers
will become sufficiently interested to delve into these subjects!

REFERENCES
[1] O. Alvarez and A. Tourin, Viscosity solutions of non-linear integro-differential
equations, Annales de l’Institut Henri Poincaré, 13 (1996), pp. 293–317.

[2] L. Andersen and J. Andreasen, Jump-diffusion models: Volatility smile fitting


and numerical methods for pricing, Rev. Derivatives Research, 4 (2000), pp. 231–262.

[3] G. Barles, R. Buckdahn, and E. Pardoux, Backward stochastic differential


equations and integral-partial differential equations, Stochastics and Stochastic Re-
ports, 60 (1997), pp. 57–83.

[4] O. Barndorff-Nielsen, Processes of normal inverse Gaussian type, Finance


Stoch., (1998), pp. 41–68.

[5] O. E. Barndorff-Nielsen and N. Shephard, Non-Gaussian Ornstein-


Uhlenbeck based models and some of their uses in financial econometrics, J. R. Statis-
tic. Soc. B, 63 (2001), pp. 167–241.

[6] D. Bates, Jumps and stochastic volatility: the exchange rate processes implicit in
Deutschemark options, Rev. Fin. Studies, 9 (1996), pp. 69–107.

[7] A. Bensoussan and J.-L. Lions, Contrôle Impulsionnel et Inéquations Quasi-


Variationnelles, Dunod, Paris, 1982.

[8] P. Carr, H. Geman, D. Madan, and M. Yor, The fine structure of asset
returns: An empirical investigation, Journal of Business, 75 (2002).

[9] P. Carr and D. Madan, Option valuation using the fast Fourier transform, J.
Comput. Finance, 2 (1998), pp. 61–73.
4
For a review, see [11, Chap. 14] and [5, 6, 30].

17
Cont et al: Option pricing models with jumps

[10] R. Cont, J.-P. Bouchaud, and M. Potters, Scaling in financial data: Sta-
ble laws and beyond, in Scale Invariance and Beyond, B. Dubrulle, F. Graner, and
D. Sornette, eds., Springer, Berlin, 1997.

[11] R. Cont and P. Tankov, Financial modelling with jump processes, Chapman &
Hall / CRC Press, 2003.

[12] R. Cont and P. Tankov, Nonparametric calibration of jump-diffusion option pric-


ing models, Journal of Computational Finance, 7 (2004), pp. 1–49.

[13] , Retrieving exponential–lévy models from option prices using relative entropy,
Rapport interne CMAP, Ecole Polytechnique, 2004.

[14] R. Cont and E. Voltchkova, Finite difference methods for option pricing in
jump-diffusion and exponential Lévy models, Rapport Interne 513, CMAP, Ecole
Polytechnique, 2003.

[15] R. Cont and E. Voltchkova, Integro-differential equations for option prices in


exponential Lévy models, Rapport Interne, CMAP, Ecole Polytechnique, 2003.

[16] M. Crandall, H. Ishii, and P. Lions, Users guide to viscosity solutions of second
order partial differential equations, Bulletin of the American Mathematical Society,
27 (1992), pp. 1–42.

[17] E. Eberlein, Applications of generalized hyperbolic Lévy motion to Finance, in


Lévy Processes—Theory and Applications, O. Barndorff-Nielsen, T. Mikosch, and
S. Resnick, eds., Birkhäuser, Boston, 2001, pp. 319–336.

[18] E. Eberlein, U. Keller, and K. Prause, New insights into smile, mispricing
and Value at Risk: The hyperbolic model, Journal of Business, 71 (1998), pp. 371–405.

[19] H. W. Engl, M. Hanke, and A. Neubauer, Regularization of Inverse Problems,


vol. 375, Kluwer Academic Publishers Group, Dordrecht, 1996.

[20] M. Garroni and J. Menaldi, Second Order Elliptic Integro-Differential Problems,


CRC Press, Boca Raton, FL, 2001.

[21] J. Jacod and A. N. Shiryaev, Limit Theorems for Stochastic Processes, Springer,
Berlin, 2nd ed., 2002.

[22] E. Jakobsen and K. Karlsen, A maximum principle for semicontinuous func-


tions applicable to integro-partial differential equations, working paper, Dept. of
Mathematics, University of Oslo, 2003.

18
Cont et al: Option pricing models with jumps

[23] I. Koponen, Analytic approach to the problem of convergence of truncated Lévy


flights towards the Gaussian stochastic process., Physical Review E, 52 (1995),
pp. 1197–1199.

[24] S. Kou, A jump-diffusion model for option pricing, Management Science, 48 (2002),
pp. 1086–1101.

[25] D. Madan, Financial modeling with discontinuous price processes, in Lévy


Processes—Theory and Applications, O. Barndorff-Nielsen, T. Mikosch, and
S. Resnick, eds., Birkhäuser, Boston, 2001.

[26] D. Madan and F. Milne, Option pricing with variance gamma martingale com-
ponents, Math. Finance, 1 (1991), pp. 39–55.

[27] A.-M. Matache, T. von Petersdorff, and C. Schwab, Fast deterministic


pricing of options on Lévy driven assets, M2AN Math. Model. Numer. Anal., 38
(2004), pp. 37–71.

[28] R. Merton, Option pricing when underlying stock returns are discontinuous, J.
Financial Economics, 3 (1976), pp. 125–144.

[29] V. Morozov, On the solution of functional equations by the method of regulariza-


tion, Soviet Math. Doklady, 7 (1966), pp. 414–417.

[30] E. Nicolato and E. Venardos, Option pricing in stochastic volatility models of


Ornstein-Uhlenbeck type, Math. Finance, 13 (2003), pp. 445–466.

[31] D. Nualart and W. Schoutens, Backward stochastic differential equations and


Feynman-Kac formula for Lévy processes, with applications in finance, Bernoulli, 7
(2001), pp. 761–776.

[32] H. Pham, Optimal stopping of controlled jump-diffusion processes: A viscosity solu-


tion approach, Journal of Mathematical Systems, 8 (1998), pp. 1–27.

[33] S. Rong, On solutions of backward stochastic differential equations with jumps and
applications, Stochastic Process. Appl., 66 (1997), pp. 209–236.

[34] K. Sato, Lévy Processes and Infinitely Divisible Distributions, Cambridge University
Press, Cambridge, UK, 1999.

[35] A. Sayah, Equations d’Hamilton Jacobi du premier ordre avec termes integro-
differentiels, Comm. Partial Differential Equations, 16 (1991), pp. 1057–1093.

[36] W. Schoutens, Lévy Processes in Finance: Pricing Financial Derivatives, Wiley,


New York, 2003.

19
Cont et al: Option pricing models with jumps

[37] H. Soner, Jump Markov Processes and Viscosity Solutions, vol. 10 of IMA Volumes
in mathematics and applications, Springer Verlag, New York, 1986, pp. 501–511.

20

You might also like